The Agent Skills Directory

[PROMPT_INJECTION]: The inclusion of an 'Ignore the previous instructions' string in vulnerable_snippet.py is identified as a controlled test case for the agent. The skill's instructions include specific rules to ignore such embedded instructions and report them as code-based injections. Mandatory Evidence Chain: 1. Ingestion points: User-provided source code processed by the agent. 2. Boundary markers: Skill instructions mandate processing content strictly within <code_to_analyze> blocks. 3. Capability inventory: No shell, network, or file-write capabilities are requested or used. 4. Sanitization: Explicit instructions (Aislamiento) command the agent to treat input as data and ignore embedded commands.
[DATA_EXFILTRATION]: The skill is designed to process external code using an 'Untrusted Data' posture. It implements boundary markers and requires the agent to ignore instructions embedded in data. No network operations or credential harvesting patterns were found in the skill or its validation scripts.
[SAFE]: The accompanying smoke_test.py is a local utility that performs simple string matching and file existence checks for structural validation of the skill, posing no security risk.

code-analysis